Reversible Decoupling Network for Single Image Reflection Removal
Hao Zhao, Mingjia Li, Qiming Hu, Xiaojie Guo
TL;DR
This paper introduces RDNet, a novel neural network architecture for single-image reflection removal that preserves valuable information through reversible encoding and dynamic feature calibration, outperforming existing methods.
Contribution
The paper proposes a reversible encoder and a transmission-rate-aware prompt generator to enhance reflection removal performance, addressing limitations of previous dual-stream networks.
Findings
RDNet outperforms state-of-the-art methods on five benchmark datasets.
Achieves top results in the NTIRE 2025 Reflection Removal Challenge.
Demonstrates superior fidelity and perceptual quality in reflection removal.
Abstract
Recent deep-learning-based approaches to single-image reflection removal have shown promising advances, primarily for two reasons: 1) the utilization of recognition-pretrained features as inputs, and 2) the design of dual-stream interaction networks. However, according to the Information Bottleneck principle, high-level semantic clues tend to be compressed or discarded during layer-by-layer propagation. Additionally, interactions in dual-stream networks follow a fixed pattern across different layers, limiting overall performance. To address these limitations, we propose a novel architecture called Reversible Decoupling Network (RDNet), which employs a reversible encoder to secure valuable information while flexibly decoupling transmission- and reflection-relevant features during the forward pass. Furthermore, we customize a transmission-rate-aware prompt generator to dynamically…
Peer Reviews
Decision·ICLR 2025 Conference Withdrawn Submission
1. The manuscript is clearly written and easy to follow. 2. The ablation study convincingly demonstrates the effectiveness of the proposed methods. 3. The proposed method achieves improved performance over existing approaches.
Novelty: The design of this network is not particularly novel. It appears to be a combination of existing methods. For instance, components such as the multi-scale reversible column encoder, hierarchical decoder, and transmission-rate-aware prompt generator have been utilized in previous works such as HRNet, and ConvNext. Computational Efficiency: Given that the network is relatively complex, there’s a lack of discussion around computational efficiency, which is important for real-time or r
1. Overall, the writing of this manuscript is good and easy to follow. 2. The ablation study can prove the effectiveness of the proposed methods. 3. The proposed method outperforms current methods.
1. The designs of this network are not very new. The proposed network looks more like a complication of existing methods. For example, the multi-scale reversible column encoder, the hierarchy decoder, and the transmission-rate-ware prompt generator have been used in existing methods (e.g., HRNet, ConvNext model). 2. The proposed network looks complex. There is a lack of comparison in FLOPs and Params in the main comparison (Table 1) and ablation study (Table 3). 3. The paper says " Our method d
1.This work presents RDNet, which incorporates a multi-column reversible encoder to effectively preserve multi-scale semantic information. By decoupling transmission and reflection features, RDNet minimizes information loss during feature interactions, significantly enhancing reflection removal accuracy. 2.The proposed transmission-rate-aware prompt generator dynamically adjusts feature representations by learning channel scaling factors from the data. This design allows RDNet to achieve strong
1.The Bidirectional Interaction Level in Figure 2 could benefit from clearer explanation, as the current description in the text is brief and may lead to misunderstandings. 2.The paper compares the proposed method with several existing reflection removal techniques. However, distinct datasets are used in the comparative experiments, which is unnecessary. Additionally, expanding the range of comparison methods would ensure a more comprehensive evaluation, as the current selection may not sufficie
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsImage Processing Techniques and Applications · Image Enhancement Techniques · Advanced Steganography and Watermarking Techniques
MethodsRDNet